AI for Green Finance

3rd place solution at Temenos Encryptcon hackathon evaluating climate projects

Temenos Encryptcon Hackathon | January 2024
Achievement: 3rd place out of 340 teams across India
Prize: $7,500 (Rs. 600,000)

Challenge

Build an AI platform to evaluate investment potential of green finance projects using Project Design Documents (PDDs), enabling investors to make data-driven decisions on climate initiatives.

Solution Architecture

1. Multimodal Document Processing

OCR-Free Parsing with Donut:

  • Processed PDDs containing text, tables, and plots without traditional OCR
  • Enabled rapid computation and semantic understanding
  • Handled diverse document layouts and formats

Document Embedding Generation:

  • Created rich representations capturing project characteristics
  • Enabled similarity-based project comparison
  • Facilitated downstream prediction tasks

2. Predictive Analytics

Carbon Credit Prediction:

  • Integrated document embeddings with time series data
  • Fine-tuned FLAN-T5 for multi-horizon forecasting
  • Predicted quantity of carbon credits generated over project lifetime
  • Forecasted carbon credit prices for upcoming years

Risk Assessment:

  • Developed heuristic comparing new projects against top performers
  • Quantified project risk based on historical success patterns
  • Provided confidence scores for investment decisions

3. Retrieval-Augmented Generation (RAG)

Intelligent Project Filtering:

  • Implemented RAG for natural language queries
  • Filtered projects based on user preferences (geography, technology, scale)
  • Presented top 3 gainers and losers based on predicted returns

Interactive Q&A:

  • Enabled users to ask questions about specific projects
  • Retrieved relevant information from PDD corpus
  • Provided evidence-based answers with source citations

Key Features

✅ Upload & Analysis: Instant evaluation of new PDDs
✅ Comparative Analytics: Benchmark against successful projects
✅ Time Series Forecasting: Predict returns over 5-10 year horizon
✅ Risk Scoring: Quantified investment risk metrics
✅ Natural Language Interface: Query projects conversationally

Technical Stack

Document Understanding: Donut (OCR-free)
Language Model: FLAN-T5
RAG Framework: LangChain
Experiment Tracking: Weights & Biases
Frontend: Gradio
Backend: Python, PyTorch, Hugging Face

Impact

  • Reduced due diligence time from days to minutes
  • Democratized access to green finance analytics
  • Enabled data-driven investment decisions in climate tech

Team Insights

“The key challenge was handling multimodal PDDs with inconsistent formats. Donut’s transformer-based approach eliminated brittle OCR pipelines, while FLAN-T5’s instruction-following capability made it ideal for integrating diverse data sources.”